Intelligent optimization method and system for lake and Hunan brocade patterns fused with style migration

Abstract

The invention provides a Hu Hunan brocade pattern intelligent optimization method and system fused with style migration, and relates to the technical field of image style migration, and the method comprises the steps: extracting common style information of a content image and a reference style image, and carrying out the guide alignment of the common style information, and obtaining a style guide feature for adjusting the pattern stylization; determining splicing control conditions of the pattern during feature splicing through the common style information and random noise features preset by pattern stylization, and generating target style information during style migration of the pattern according to the splicing control conditions and the style guide features, performing element filling on an image region missing during style migration according to the target style information to obtain a style migration image; and performing image reconstruction according to the detail features of the style migration image and the content image, and generating an optimized Hu-Hunan brocade pattern. According to the method, on the basis that the original pattern structure of the Hu Xiang brocade is completely reserved, adaptive migration of the full-format style of the pattern can be achieved.

Description

Technical Field

[0001] This application relates to the field of image style transfer technology, and more specifically, to a method and system for intelligent optimization of Hunan brocade patterns that incorporates style transfer. Background Technology

[0002] Image style transfer refers to an image processing technique that, while preserving the core content information such as the structure and element layout of the pattern to be optimized, accurately transfers the style features carried by the reference style image to the pattern to be optimized through feature extraction, style alignment, and adaptive fusion. Its core principle is to decouple the content features and style features of the image through algorithms, and then re-fuse them according to specific style constraint rules. This allows the optimized pattern to not only fully retain the core creativity and structural framework of the original design, but also possess the unified traditional style attributes that conform to the intangible cultural heritage standards, providing efficient style standardization technical support for pattern innovation design and industrial production.

[0003] However, the stylization optimization of traditional Hunan brocade patterns lacks an optimization style constraint mechanism that allows for spatial adaptive integration of the core pattern structure and traditional style characteristics. This leads to issues such as awkward filling of missing areas during the optimization process, making it difficult to maintain the integrity of both the stylized and original patterns. Therefore, how to achieve full-frame style adaptation and transfer of patterns while completely preserving the original pattern structure of Hunan brocade is a challenge facing the industry. Summary of the Invention

[0004] This application provides a method and system for intelligent optimization of Hunan brocade patterns that integrates style transfer, which can achieve full-frame style adaptation and transfer of patterns while completely preserving the original pattern structure of Hunan brocade.

[0005] Firstly, this application provides an intelligent optimization method for Hunan brocade patterns that incorporates style transfer, the optimization method comprising the following steps:

[0006] Obtain the content image and reference style image of the Hunan brocade pattern to be optimized;

[0007] Extract common style information from the content image and the reference style image, and perform guided alignment on the common style information to obtain style guidance features for adjusting pattern stylization;

[0008] The splicing control conditions of the pattern during feature splicing are determined by the common style information and the random noise features preset by the pattern stylization. The target style information for style transfer of the pattern is generated according to the splicing control conditions and the style guidance features. Then, the missing image areas during style transfer are filled with elements by the target style information to obtain the style-transferred image.

[0009] Image reconstruction is performed based on the detailed features of the style transfer image and the content image to generate an optimized Hunan brocade pattern.

[0010] In this embodiment, extracting the common style information of the content image and the reference style image specifically includes:

[0011] The content image and the reference style image are decomposed into mutually independent structure-style distribution constraints by using a joint graph;

[0012] Based on the structure-style distribution constraints, the common primitives of the content image and the reference style image are determined;

[0013] The common primitives are injected into the structural phase and style magnitude corresponding to the content image and the reference style image, respectively, to obtain common style information.

[0014] In this embodiment, guiding the alignment of the common style information to obtain style guidance features for adjusting pattern stylization specifically includes:

[0015] Construct a dense field mapping relationship between the common style information and the target pattern space, and determine the style distribution weight of each pixel;

[0016] The initial distribution vector in the spatiotemporal domain of the adjusted pattern stylization is determined based on the style distribution weights and the local structure tensor of the target pattern.

[0017] The initial distribution vector is optimized for edge preservation to generate a spatially consistent alignment field;

[0018] The spatial consistency alignment field is fused layer by layer with the multi-scale feature map of the target pattern to obtain style guidance features that adjust the pattern stylization.

[0019] In this embodiment, determining the initial distribution vector in the spatiotemporal domain for adjusting pattern stylization based on the style distribution weights and the local structure tensor of the target pattern specifically includes:

[0020] Based on the local structure tensor, feature decomposition is performed to extract the principal direction vector and anisotropic intensity coefficient of the target pattern at the pixel point;

[0021] The principal direction vector is weighted and fused according to the style distribution weight and the anisotropy intensity coefficient to generate a local style guidance vector at each pixel.

[0022] The local style-guided vector is regularized to obtain the initial distribution vector in the spatial domain when adjusting the pattern stylization.

[0023] In this embodiment, determining the splicing control conditions for pattern feature splicing based on the common style information and the pre-defined random noise features of pattern stylization specifically includes:

[0024] A dual-modal hybrid sampling distribution is constructed in the joint feature space based on the probability distribution of the common style information and the random noise features;

[0025] The spatial adaptive control coefficients of each feature channel are generated during feature stitching using the dual-modal mixed sampling distribution and the preset learnable stitching weight mapping function.

[0026] Based on the spatial adaptive control coefficients and the local correlation metric between feature maps, the stitching control conditions for the pattern during feature stitching are determined.

[0027] In this embodiment, generating target style information for style transfer of the pattern based on the splicing control conditions and the style guidance features specifically includes:

[0028] Construct a multi-scale distribution mapping relationship in the joint feature space based on the splicing control conditions and the style guidance features;

[0029] Based on the multi-scale distribution mapping relationship and the spatial location encoding of the target pattern, the dynamic modulation coefficients of style information in each spatial region are determined;

[0030] The style guidance features are spatially adaptively fused using the dynamic modulation coefficients to generate a multi-channel feature map of preliminary target style information.

[0031] The multi-channel feature map is regularized iteratively based on the boundary continuity constraint in the splicing control conditions to obtain the target style information when performing style transfer on the pattern.

[0032] In this embodiment, determining the dynamic modulation coefficients of style information in each spatial region based on the multi-scale distribution mapping relationship and the spatial location encoding of the target pattern specifically includes:

[0033] Based on the multi-scale distribution mapping relationship and the spatial location encoding, a conditional probability density distribution function is constructed in the joint feature space;

[0034] The preliminary modulation weights of the regional style response are calculated by integrating the conditional probability density distribution function over the spatial regions of the target pattern.

[0035] Adaptive regularization is performed based on the initial modulation weights and neighborhood space continuity constraints to obtain the dynamic modulation coefficients of style information in each spatial region.

[0036] In this embodiment, the process of filling in missing image regions during style transfer with the target style information to obtain a style-transferred image specifically includes:

[0037] Based on the target style information and the known pixels in the neighborhood of the missing image region during style transfer, determine the style guidance distribution of each pixel in the missing region;

[0038] The set of confidence elements within the missing region is determined by the style-guided distribution;

[0039] The set of confidence elements is subjected to style consistency processing to obtain a style transfer image.

[0040] In this embodiment, image reconstruction based on the detailed features of the style transfer image and the content image to generate the optimized Hunan brocade pattern specifically includes:

[0041] The texture complementarity coefficients for image reconstruction are determined based on the detail feature maps of the style transfer image and the content image.

[0042] Feature-level fusion is performed using the texture complementarity coefficient and multi-scale structural similarity constraints to generate an intermediate feature map with consistent texture of Hunan brocade.

[0043] The optimized Hunan brocade pattern is obtained by adaptive reconstruction based on the intermediate feature map and the preset Hunan brocade color style dictionary.

[0044] Secondly, this application provides an intelligent optimization system for Hunan brocade patterns that integrates style transfer, used to execute an intelligent optimization method for Hunan brocade patterns that integrates style transfer. The optimization system includes:

[0045] The acquisition module is used to acquire the content image and reference style image of the Hunan brocade pattern to be optimized;

[0046] The guided alignment module is used to extract common style information between the content image and the reference style image, and to perform guided alignment on the common style information to obtain style guidance features for adjusting the pattern stylization.

[0047] The element filling module is used to determine the splicing control conditions of the pattern when splicing features by using the common style information and the random noise features preset by the pattern stylization, generate target style information for style transfer of the pattern according to the splicing control conditions and the style guidance features, and then fill in the missing image areas during style transfer with the target style information to obtain the style transfer image.

[0048] The image reconstruction module is used to reconstruct the image based on the detailed features of the style transfer image and the content image, and generate an optimized Hunan brocade pattern.

[0049] The technical solutions provided by the embodiments disclosed in this application have the following beneficial effects:

[0050] Obtain the content image and reference style image of the Hunan brocade pattern to be optimized; extract the common style information of the content image and the reference style image, and guide the alignment of the common style information to obtain style guidance features for adjusting the pattern stylization; determine the splicing control conditions of the pattern during feature splicing through the common style information and the random noise features preset for pattern stylization; generate target style information for style transfer of the pattern based on the splicing control conditions and the style guidance features; then fill in the missing image areas during style transfer with the target style information to obtain a style transfer image; perform image reconstruction based on the style transfer image and the detailed features of the content image to generate the optimized Hunan brocade pattern.

[0051] Therefore, in this application, image reconstruction is performed based on the detailed features of the style transfer image and the content image to generate an optimized Hunan brocade pattern. Specifically, determining the style guiding feature provides a directional style adjustment benchmark adapted to the multi-scale structural characteristics of Hunan brocade, thereby achieving hierarchical and spatially differentiated precise control over the full-page stylization process of the pattern. This style guiding feature anchors the core traditional style norms of Hunan brocade's intangible cultural heritage attributes, establishes a spatial correspondence between style features and the original pattern structure, and allows for setting differentiated style constraint strengths for the core pattern area and flat texture areas. While protecting the integrity of the original pattern structure, it ensures the adaptability and consistency of the full-page style features, thus providing a basis for the full-page style of the pattern. The precise adaptation and migration provides a stable and controllable basis; by determining the style transfer image, a full-width brocade image carrier that combines the integrity of the original pattern structure with the uniformity of the target style can be obtained, thus completing the entire link of Hunan brocade pattern from feature domain style constraints to pixel domain complete presentation. The style transfer image uses the target style information as a global constraint, completely continuing the core structural information such as the pattern layout and line outline of the original brocade, realizing a smooth transition and unified adaptation of the full-width traditional style, while repairing the details loss and texture breakage problems that occurred during the style transfer process, ensuring the consistency of the pattern style and the adaptability of the weaving process, and finally achieving the goal of completely preserving the original pattern structure and adapting and migrating the traditional style to the full-width image.

[0052] In summary, the technical solution adopted in this application can achieve full-frame style adaptation and transfer of patterns while completely preserving the original pattern structure of Hunan brocade. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this embodiment of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0054] Figure 1 This is an exemplary flowchart of an intelligent optimization method for Hunan brocade patterns that integrates style transfer, provided in this application.

[0055] Figure 2 This is a flowchart illustrating the process of determining style guidance features based on the information provided in this application;

[0056] Figure 3 This is a flowchart illustrating the process of generating target style information based on the information provided in this application;

[0057] Figure 4 This is a schematic diagram of the multi-scale feature extraction structure of Hunan brocade patterns provided in this application;

[0058] Figure 5 This is a module structure diagram of an intelligent optimization system for Hunan brocade patterns that integrates style transfer, provided in this application. Detailed Implementation

[0059] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0060] This application provides an intelligent optimization method and system for Hunan brocade patterns using style transfer. The core of this method involves acquiring a content image and a reference style image of the Hunan brocade pattern to be optimized; extracting common style information from the content image and the reference style image; guiding and aligning the common style information to obtain style guidance features for adjusting the pattern's stylization; determining splicing control conditions for the pattern during feature splicing based on the common style information and preset random noise features for pattern stylization; generating target style information for style transfer of the pattern based on the splicing control conditions and the style guidance features; filling in missing image regions during style transfer with the target style information to obtain a style-transferred image; and reconstructing the image based on the style-transferred image and the detailed features of the content image to generate the optimized Hunan brocade pattern.

[0061] Example 1: To better understand the above technical solution, the following will provide a detailed description of the technical solution in conjunction with the accompanying drawings and specific implementation methods. (Refer to...) Figure 1 As shown in the figure, this is an exemplary flowchart of an intelligent optimization method for Hunan brocade patterns that integrates style transfer, according to this embodiment of the present application. The optimization method includes the following steps:

[0062] In step S1, the content image and reference style image of the Hunan brocade pattern to be optimized are obtained.

[0063] In practice, a flatbed scanner with an optical resolution of 600 dpi or higher can be used to scan the entire surface of the Hunan brocade to be optimized, or the original digital file of the brocade to be optimized can be directly used to obtain the original digital image of the brocade. The original digital image is then subjected to median filtering for noise reduction, white balance color shift correction, and edge cropping to remove redundancy, resulting in a content image that retains only the complete brocade pattern and is free of interference information. At the same time, standard brocade images of the same theme and category as the pattern to be optimized are selected from the digital resource database of Hunan brocade in the museum and the standard brocade sample database certified by intangible cultural heritage institutions as style reference sources. The selected images undergo a preprocessing process that is completely consistent with the content image to obtain a reference style image that matches the aspect ratio and resolution of the content image.

[0064] It should be noted that, in this application, the content image refers to the pattern and layout information that needs to be fully preserved in the Hunan brocade to be optimized; the reference style image refers to the standard brocade image that provides a target style reference for the optimization of Hunan brocade.

[0065] In step S2, common style information of the content image and the reference style image is extracted, and the common style information is guided and aligned to obtain style guidance features for adjusting the pattern stylization.

[0066] In this embodiment, extracting the common style information between the content image and the reference style image can be achieved through the following steps:

[0067] The content image and the reference style image are decomposed into mutually independent structure-style distribution constraints by using a joint graph;

[0068] Based on the structure-style distribution constraints, the common primitives of the content image and the reference style image are determined;

[0069] The common primitives are injected into the structural phase and style magnitude corresponding to the content image and the reference style image, respectively, to obtain common style information.

[0070] In practice, firstly, two-dimensional discrete Fourier transforms are performed on the preprocessed content image and the reference style image to transform the image from the spatial domain to the frequency domain, resulting in two sets of corresponding frequency domain complex matrices. A dual-input joint spectral analysis framework is constructed as a joint spectrum. Based on the decoupling of the phase component corresponding to the structural information and the amplitude component corresponding to the style information in the frequency domain, joint spectral decomposition is performed on the two sets of complex matrices to output the structural distribution constraints and style distribution constraints corresponding to the content image and the reference style image, respectively. These are the structure-style distribution constraints, which are independent of each other and have no feature intersections, thus achieving complete decoupling of structure and style. Then, based on the structure-style distribution constraints, the style distribution constraint intervals of the content image and the reference style image are locked respectively. The feature dimensions corresponding to the structure distribution constraints are excluded, and primitives are extracted only within the style feature dimension. A sparse coding algorithm can be used to perform overcomplete dictionary learning on the frequency domain features within the style distribution constraints of the two images respectively, resulting in two sets of corresponding style primitive dictionaries. Cosine similarity matching is performed on the primitives in the two sets of dictionaries, and primitives with similarity in the threshold range of 0.7-0.9 are selected. The threshold range can be set according to the frequency features of the image style distribution constraints, which is not limited here, and is used as the common primitives of the two images. Finally, the structural phases corresponding to the content image and the reference style image are preserved to ensure that the pattern structure remains unchanged. At the same time, based on the structure-style distribution constraint, the style amplitude ranges corresponding to the two images are locked. The common primitives obtained by screening are mapped to the frequency domain space of the two images respectively. With the common primitives as the core weights, the amplitude components in the corresponding style amplitude ranges are weighted and updated to complete the injection of common primitives. The updated frequency domain complex matrix is ​​subjected to a two-dimensional discrete Fourier inverse transform to convert it back to the spatial domain to obtain the common style information.

[0071] It should be noted that, in this application, the joint spectrum refers to a dual-input joint spectrum analysis tool that performs structural and stylistic decoupling decomposition on the content image and the reference style image; the structure-style distribution constraint refers to the independent distribution range of the image structure information and style information, respectively; the common primitive is the smallest feature unit that represents the core style attributes of Hunan brocade shared by the content image and the reference style image; the structural phase refers to the content information that carries the spatial position, outline structure and element layout of the image pattern; the style amplitude refers to the style information that carries the color scheme, texture and weaving texture of the image; and the common style information represents the style features of Hunan brocade shared by the content image and the reference style image.

[0072] Preferably, in this embodiment, the common style information is guided and aligned to obtain style guidance features for adjusting the stylization of the pattern, with reference to... Figure 2 As shown in the figure, this is a flowchart illustrating the process of determining style guidance features in some embodiments of this application. In this embodiment, determining style guidance features can be achieved through the following steps:

[0073] In step S21, a dense field mapping relationship between the common style information and the target pattern space is constructed, and the style distribution weight of each pixel is determined;

[0074] In step S22, the initial distribution vector in the spatial domain of the adjusted pattern stylization is determined based on the style distribution weights and the local structure tensor of the target pattern.

[0075] In step S23, the initial distribution vector is optimized to preserve the edges, generating a spatially consistent alignment field;

[0076] In step S24, the spatial consistency alignment field is fused layer by layer with the multi-scale feature map of the target pattern to obtain style guidance features for adjusting the pattern stylization.

[0077] In specific implementation, firstly, the preprocessed common style information and the target pattern space are normalized to a unified numerical range of 0-1. A bilinear interpolation full-pixel mapping model is constructed as a dense field mapping relationship, realizing a one-to-one correspondence between each feature dimension of the common style information and each pixel in the target pattern space. Then, the fit degree of each pixel to the common style information is calculated based on the local variance of the texture in the 3×3 neighborhood of each pixel in the target pattern. After normalizing the fit degree value, the style distribution weights corresponding to each pixel in the target pattern space are obtained. Next, Gaussian smoothing is performed on the single-channel grayscale image of the target pattern, and the image gradient components in the horizontal and vertical directions of each pixel are calculated. Based on the gradient components, a 2×2-dimensional local structure tensor is constructed for each pixel. The elements of the local structure tensor are the product of the gradient components and the Gaussian weighted sum of the neighborhood. Then, the style distribution weights of each pixel are used as vector magnitude constraints, and the principal eigenvectors of the local structure tensor are used as vector direction constraints to complete the vector synthesis of each pixel, obtaining the initial distribution vector in the spatial domain. Then, based on the eigenvalue differences of the local structure tensor of the target pattern, the edge regions and non-edge flat regions of the target pattern are identified. High structure constraint weights are set for the edge regions and low constraint weights for the non-edge regions. A total variation minimization algorithm is used, with the initial distribution vector as the optimization benchmark and the structural constraints of the edge regions as the regularization term, to iteratively optimize the initial distribution vector of the entire image, eliminating local vector mutations. After iterative convergence, a spatial consistency alignment field is obtained. Finally, a 19-layer convolutional neural network for visual geometry pre-trained on the ImageNet dataset is used. The fully connected layers and classification layers at the end of the network are removed, and the outputs of five convolutional layers of different depths are selected to construct multi-scale feature maps of the target pattern from high resolution to low resolution. The spatial consistency alignment field is downsampled to a size that matches the feature maps of each layer and used as a fusion weight to weight and fuse the feature maps of the corresponding layers. Then, the fused features of each layer are concatenated to obtain the style guidance features that adjust the stylization of the pattern.

[0078] It should be noted that, in this application, guided alignment refers to the process of mapping common style information to the feature space of the target pattern, eliminating the spatial distribution deviation between style features and the target pattern; the target pattern space refers to the pixel-feature domain of the target pattern that carries the complete pixel distribution and feature dimensions of the Hunan brocade pattern to be optimized; dense field mapping relationship refers to the model that establishes a one-to-one correspondence between each feature dimension of the common style information and each pixel in the target pattern space; style distribution weight is a characterization of the degree of adaptation of each pixel in the target pattern space to the common style information; local structure tensor is a characterization of the local structural properties of the pattern edge direction, texture gradient, and structural distribution in the local area of ​​the target pattern; initial distribution vector is a characterization of the initial distribution direction and intensity of style features in the target pattern space domain; edge-preserving alignment optimization refers to... While optimizing the style distribution vector, the pattern edge structure of the target pattern is fully preserved, avoiding the spatial vector optimization process of edge blurring and deformation during stylization; the spatial consistency alignment field refers to the pixel-level style mapping field that ensures the continuous distribution and smooth transition of style features across the entire target pattern space, without local style abrupt changes; the multi-scale feature map is a set of multi-resolution feature maps that characterize the structural and textural features of the target pattern at different resolution levels; layer-by-layer fusion refers to the feature processing process that, based on the hierarchical structure of the multi-scale feature map of the target pattern, matches the spatial consistency alignment field according to the feature size of the corresponding level, performs independent feature fusion with the feature map of each level, and then integrates the features fused from each level in an orderly manner; style-guided features refer to the feature set of stylization transfer of Hunan brocade patterns.

[0079] Furthermore, in this embodiment, determining the initial distribution vector in the spatiotemporal domain of the adjusted pattern stylization based on the style distribution weights and the local structure tensor of the target pattern can be achieved through the following steps:

[0080] Based on the local structure tensor, feature decomposition is performed to extract the principal direction vector and anisotropic intensity coefficient of the target pattern at the pixel point;

[0081] The principal direction vector is weighted and fused according to the style distribution weight and the anisotropy intensity coefficient to generate a local style guidance vector at each pixel.

[0082] The local style-guided vector is regularized to obtain the initial distribution vector in the spatial domain when adjusting the pattern stylization.

[0083] In practice, firstly, for each pixel of the target pattern, the 2×2 local structure tensor is subjected to real symmetric matrix eigenvalue decomposition using the well-known Jacobi iteration method. This yields two non-negative eigenvalues ​​and their corresponding orthogonal unit eigenvectors. The unit eigenvector corresponding to the larger eigenvalue is determined as the principal direction vector at that pixel, accurately corresponding to the edge and line direction of the brocade pattern. Then, based on the ratio of the difference to the sum of the two eigenvalues, the anisotropy intensity coefficient of that pixel is calculated. The anisotropy intensity coefficient ranges from 0 to 1, representing the strength of the directionality of the local structure. Next, the style distribution weight corresponding to each pixel of the target pattern is multiplied by the anisotropy intensity coefficient to obtain the vector magnitude constraint coefficient for that pixel. Using the principal direction vector of that pixel as a fixed direction reference and the vector magnitude constraint coefficient as the vector magnitude, a weighted synthesis of single-pixel vectors is performed. The synthesized vectors are then subjected to direction normalization and global magnitude normalization to eliminate numerical magnitude differences between different pixels, ultimately generating independent local style guidance vectors for each pixel. Finally, a 5×5 square neighborhood window is constructed with each pixel of the target pattern as the center, covering the local style guidance vectors of all adjacent pixels within the window. A Gaussian weighted vector averaging method is used, with Gaussian weights assigned based on the Euclidean distance between the neighboring pixels and the center pixel. The vectors in the neighborhood are then weighted and averaged to obtain the regularized vector of the center pixel. This process is repeated for all pixels across the entire target pattern to complete the processing, ultimately yielding the initial distribution vector in the spatial domain.

[0084] It should be noted that, in this application, eigenvalue decomposition refers to a linear algebraic processing method that performs standardized matrix decomposition on the real symmetric local structure tensor corresponding to each pixel of the target pattern, decoupling the eigenvector representing the direction of the brocade pattern structure from the eigenvalue representing the strength of the structural directionality; the principal direction vector represents the direction of the edge and texture of the brocade pattern at the pixel of the target pattern; the anisotropy intensity coefficient represents the strength of the directionality of the local brocade structure at the pixel of the target pattern; the local style guidance vector represents the distribution direction and adaptation strength of the style features at a single pixel; neighborhood consistency regularization refers to the regularization process that constrains the spatial continuity of the style guidance vectors of adjacent pixels, eliminates abrupt changes in local vectors, and ensures a smooth transition of style distribution; the initial distribution vector represents the initial distribution direction and intensity of the style features in the entire spatial domain of the target pattern.

[0085] In step S3, the stitching control conditions for the pattern during feature stitching are determined by the common style information and the preset random noise features of the pattern stylization. Based on the stitching control conditions and the style guidance features, target style information is generated for style transfer of the pattern. Then, the target style information is used to fill in the missing image regions during style transfer, resulting in a style-transferred image.

[0086] In this embodiment, determining the splicing control conditions for the pattern during feature splicing using the common style information and the preset random noise features of the pattern stylization can be achieved through the following steps:

[0087] A dual-modal hybrid sampling distribution is constructed in the joint feature space based on the probability distribution of the common style information and the random noise features;

[0088] The spatial adaptive control coefficients of each feature channel are generated during feature stitching using the dual-modal mixed sampling distribution and the preset learnable stitching weight mapping function.

[0089] Based on the spatial adaptive control coefficients and the local correlation metric between feature maps, the stitching control conditions for the pattern during feature stitching are determined.

[0090] In specific implementation, firstly, the preprocessed common style information and the pre-defined random noise features of the pattern stylization are linearly projected onto a joint feature space with perfectly matched dimensions to eliminate the dimensional differences between the two. A kernel density estimation algorithm is then used to fit the style mode probability distribution corresponding to the common style information and the noise mode probability distribution corresponding to the random noise features, respectively. Then, the two single-modal distributions are weighted and superimposed using fixed mixing weights to construct a bimodal mixed sampling distribution within the joint feature space. Next, uniform sampling is performed at equal intervals from the bimodal mixed sampling distribution to obtain a sampling feature set covering the full distribution range of both style and noise modes. This sampling feature set is then input into a pre-defined learnable splicing weight mapping function, which is a 3-layer fully connected network pre-trained based on the stylized sample set of Hunan brocade from the museum's collection. The function outputs weight values ​​that correspond one-to-one with the feature channels and spatial locations. Finally, the weight values ​​are normalized to the 0-1 range to generate spatial adaptive control coefficients for each feature channel. Finally, the local correlation metric between each pair of the content feature map, style feature map, and noise feature map to be stitched is calculated. The cosine similarity of the 5×5 neighborhood window is used to calculate the feature correlation value of each local region. The spatial adaptive control coefficient is used as the weight constraint, and the local correlation metric is used as the boundary smoothing constraint to determine the channel fusion weight rule, spatial region division rule, and stitching boundary transition rule for feature stitching. The three types of rules together constitute the stitching control conditions of the pattern during feature stitching.

[0091] It should be noted that, in this application, random noise features refer to a set of standardized Gaussian random features that provide controllable random perturbations for feature splicing, enrich the weaving texture details of Hunan brocade, buffer the boundary transition of feature splicing, and avoid abrupt breaks in the splicing; feature splicing refers to a feature processing mechanism that constrains the pattern structure and unifies the traditional style; probability distribution is a representation of the distribution law and aggregation characteristics of feature values; joint feature space refers to a standardized feature dimension space that carries common style information and random noise features; bimodal mixed sampling distribution is a representation of the style distribution characteristics of common style information and the perturbation distribution characteristics of random noise features; learnable splicing weight mapping function is a mathematical model that maps probability distribution features to feature channel-level splicing weights; feature channel refers to a feature dimension unit that independently carries a single-dimensional specific feature of Hunan brocade; spatial adaptive control coefficient refers to the set of fusion weights for different spatial positions and different feature channels when controlling feature splicing; local correlation metric is a quantified value of local feature similarity that represents the feature correlation of local areas of the feature map to be spliced; splicing control conditions refer to the transition rules that constrain the fusion weights of each feature and the division of splicing regions during the feature splicing process.

[0092] Preferably, in this embodiment, target style information for style transfer of the pattern is generated based on the splicing control conditions and the style guidance features, with reference to... Figure 3 As shown in the figure, this is a schematic diagram of the process for generating target style information in some embodiments of this application. In this embodiment, the generation of target style information can be achieved by the following steps:

[0093] In step S31, a multi-scale distribution mapping relationship in the joint feature space is constructed based on the splicing control conditions and the style guidance features;

[0094] In step S32, the dynamic modulation coefficients of style information in each spatial region are determined based on the multi-scale distribution mapping relationship and the spatial location encoding of the target pattern;

[0095] In step S33, the style guidance features are spatially adaptively fused using the dynamic modulation coefficients to generate a multi-channel feature map of preliminary target style information;

[0096] In step S34, the multi-channel feature map is regularized iterated according to the boundary continuity constraint in the splicing control conditions to obtain the target style information when performing style transfer on the pattern.

[0097] In specific implementation, firstly, the channel fusion weight rules and spatial region division rules in the splicing control conditions, along with the multi-scale hierarchical features of the style-guided features, are linearly projected onto a joint feature space with perfectly matched dimensions, eliminating the dimensional differences between the two types of features. Based on the five standard feature layers of a 19-layer convolutional neural network for visual geometry pre-trained on the ImageNet dataset, a bilinear mapping model between style features and the target pattern space is constructed for each resolution level. Integrating all hierarchical mapping models yields a multi-scale distribution mapping relationship within the joint feature space. Next, a sine-cosine positional encoding method is used to generate corresponding multi-scale spatial positional codes for each spatial region of the target pattern. These codes include the absolute positional attributes and pattern structure type attributes of the region. The spatial positional codes are input into the multi-scale distribution mapping relationship, outputting the corresponding style distribution baseline weights for each region. Then, combined with the region weight rules in the splicing control conditions, the baseline weights are normalized in the 0-1 range, ultimately obtaining the dynamic modulation coefficients corresponding to each spatial region. Then, the dynamic modulation coefficients are downsampled to a size that perfectly matches the feature maps of each level of the style guidance feature using bilinear interpolation, resulting in hierarchical modulation coefficients corresponding one-to-one with each feature level. For each feature level, the hierarchical modulation coefficients are used as weights and fused with the style guidance feature map of that level channel by channel and pixel by pixel to obtain the fused style feature maps of each level. All the fused feature maps of all levels are then sequentially stitched together according to the channel dimension to generate a multi-channel feature map of the preliminary target style information. Finally, based on the boundary continuity constraint in the stitching control condition, the stitching boundary positions of different spatial regions in the multi-channel feature map are identified. Using the cosine similarity of the features on both sides of the boundary as the optimization objective, a total variational regularization loss function is constructed. Gradient descent is used to iteratively optimize the multi-channel feature map to eliminate feature abrupt changes at the boundary. After iterative convergence, the optimized feature map is mapped back to the image pixel space through deconvolution, finally obtaining the target style information for style transfer of the pattern.

[0098] It should be noted that, in this application, the multi-scale distribution mapping relationship refers to the set of multi-resolution mapping models that establish a one-to-one correspondence between style guidance features and the target pattern space at different resolution levels; spatial position encoding is the representation of the absolute position attribute and pattern structure type of each spatial region of the target pattern; dynamic modulation coefficient is the control of the injection intensity and distribution weight of style information in different spatial regions of the target pattern; spatial adaptive fusion is the feature fusion processing mechanism that performs feature fusion on the pattern structure attributes of different spatial regions of the Hunan brocade target pattern and dynamically adjusts the fusion weight of style guidance features in the corresponding regions; the preliminary target style information channel feature map refers to the multi-resolution style feature map that carries the multi-dimensional style features after spatial adaptive fusion; boundary continuity constraint is the regularization constraint rule that constrains the feature continuity of style features at the splicing boundary of different regions of the pattern; regularization iteration is the iterative optimization process that uses the boundary continuity constraint in the splicing control condition as the criterion, and through a preset regularization loss function and iterative optimization algorithm, iteratively optimizes and corrects the initially generated multi-channel style feature map to eliminate local feature mutations and ensure the continuity of the style space of the whole image; target style information refers to the set of pixel-level style features covering all regions of the pattern when the full-width style of the Hunan brocade pattern is transferred.

[0099] Furthermore, in this embodiment, determining the dynamic modulation coefficients of style information in each spatial region based on the multi-scale distribution mapping relationship and the spatial location encoding of the target pattern can be achieved using the following steps:

[0100] Based on the multi-scale distribution mapping relationship and the spatial location encoding, a conditional probability density distribution function is constructed in the joint feature space;

[0101] The preliminary modulation weights of the regional style response are calculated by integrating the conditional probability density distribution function over the spatial regions of the target pattern.

[0102] Adaptive regularization is performed based on the initial modulation weights and neighborhood space continuity constraints to obtain the dynamic modulation coefficients of style information in each spatial region.

[0103] In practice, firstly, the multi-level style features output from the multi-scale distribution mapping relationship are mapped to the spatial location codes of the corresponding spatial regions of the target pattern through linear projection, eliminating the dimensional differences between the two types of features. A multivariate Gaussian kernel density estimation algorithm is then used, with the spatial location codes as conditional variables and the multi-scale style features as random variables, to fit the correspondence between conditional probabilities and feature values, ultimately constructing the conditional probability density distribution function within the joint feature space. Next, the target pattern is divided into equally sized grids to obtain non-overlapping independent spatial regions. A corresponding spatial location code is matched to each region. For each independent spatial region, the corresponding location code is substituted into the conditional probability density distribution function, and the Gauss-Legendal numerical integration method is used to solve for the style feature dimension by definite integral, obtaining the style response probability value for that region. The probability values ​​for the entire region are then normalized to the 0-1 interval to obtain the preliminary modulation weights of the regional style response. Finally, a 3×3 square neighborhood window is constructed with each spatial region of the target pattern as the center. Based on the Euclidean distance between each neighborhood region and the central region, and the similarity of the pattern structure, a weight matrix constraining the spatial continuity of the neighborhood is constructed. Using the initial modulation weights as the initial values ​​for optimization, a total variational regularization method is adopted, with the neighborhood weight matrix as the regularization term constraint, to perform iterative smooth optimization, eliminating abrupt weight changes in adjacent regions. After iterative convergence, the dynamic modulation coefficients of style information in each spatial region are obtained.

[0104] It should be noted that in this application, the conditional probability density distribution function is a probability distribution of style features conditioned on the spatial location of the target pattern; the marginalization integral is a mathematical operation method for decoupling the style response probability of a single spatial region from the joint conditional probability distribution; the initial modulation weight is an initial benchmark value representing the style injection intensity of each spatial region of the target pattern; the neighborhood spatial continuity constraint is a regularization constraint rule that constrains the smooth transition of the modulation weights of adjacent spatial regions of the target pattern, avoiding local jumps during the stylization process; the adaptive regularization is a regularization optimization mechanism that takes the initial modulation weights of each spatial region of the target pattern as the optimization object, takes the neighborhood spatial continuity constraint as the core criterion, and dynamically adjusts the regularization constraint intensity according to the pattern structure attributes of different regions of Hunan brocade, eliminating weight abrupt changes in adjacent regions while preserving the differences in regional style modulation weights, and generating dynamic modulation coefficients that combine regional adaptability and spatial continuity; the dynamic modulation coefficient is a set of regional weight coefficients that control the style information injection intensity of each spatial region of the target pattern.

[0105] In this embodiment, the missing image regions during style transfer are filled with elements using the target style information to obtain a style-transferred image, which can be achieved through the following steps:

[0106] Based on the target style information and the known pixels in the neighborhood of the missing image region during style transfer, determine the style guidance distribution of each pixel in the missing region;

[0107] The set of confidence elements within the missing region is determined by the style-guided distribution;

[0108] The set of confidence elements is subjected to style consistency processing to obtain a style transfer image.

[0109] In specific implementation, firstly, image semantic segmentation methods are used to accurately identify the complete boundaries and pixel coordinate range of the image region missing during style transfer. Known pixels within an 8-neighborhood of the missing region are extracted to obtain their pattern structure features and corresponding target style information. A multivariate Gaussian kernel density estimation algorithm is used, with the style features of the known neighboring pixels as fitting samples and the target style information as a global constraint, to fit the style-guided distribution corresponding to each pixel in the missing region. Then, the missing region is divided into equal-sized grids, with the grid unit size perfectly matching the minimum weaving unit size of Hunan brocade, ensuring that the filling elements conform to the weaving rules of brocade. Using the style-guided distribution as a probability constraint, a Markov chain Monte Carlo sampling method is used to iteratively sample each grid unit multiple times, generating candidate brocade pattern elements. Elements with a fit higher than a preset threshold to the style-guided distribution are selected and integrated to obtain a set of confidence elements within the missing region. Finally, the set of confidence elements is arranged in an orderly manner according to the grid position of the missing area to complete the preliminary filling of the pattern structure of the missing area. The Poisson image fusion method is used to smoothly fuse the element splicing boundary with the known pixels in the neighborhood as the boundary reference to eliminate boundary color difference and structural discontinuity. Then, with the target style information as the global reference, histogram matching and texture consistency calibration are performed on the filled area and the whole image to complete the style consistency processing and finally obtain the style transfer image.

[0110] It should be noted that in this application, element filling refers to the standardized image processing process for repairing missing areas; the missing image area refers to the image area where pattern details are lost, textures are broken, and effective information is missing during the style transfer process; the known neighboring pixels refer to the effective pixels located around the missing area that retain the complete brocade pattern structure and style information; the style-guided distribution is a statistical distribution model that characterizes the probability of style feature values ​​of each pixel in the missing area and constrains the style direction and distribution law of the filling elements; the confidence element set refers to the set of high-confidence Hunan brocade pattern elements that conform to the style-guided distribution and the constraints of the neighboring pattern structure; style consistency processing refers to the image processing process of performing boundary smoothing fusion and global style calibration on the confidence element set; the style transfer image refers to the complete image that completes full-area style transfer and complete filling of missing areas, while retaining the original brocade core pattern and the target traditional style attributes.

[0111] In step S4, image reconstruction is performed based on the detailed features of the style transfer image and the content image to generate an optimized Hunan brocade pattern.

[0112] In this embodiment, the optimized Hunan brocade pattern can be generated by reconstructing the image based on the detailed features of the style transfer image and the content image using the following steps:

[0113] The texture complementarity coefficients for image reconstruction are determined based on the detail feature maps of the style transfer image and the content image.

[0114] Feature-level fusion is performed using the texture complementarity coefficient and multi-scale structural similarity constraints to generate an intermediate feature map with consistent texture of Hunan brocade.

[0115] The optimized Hunan brocade pattern is obtained by adaptive reconstruction based on the intermediate feature map and the preset Hunan brocade color style dictionary.

[0116] In specific implementation, firstly, the Laplacian pyramid decomposition method is used to decompose the style transfer image and content image into 5 levels, extracting the high-frequency detail components corresponding to each level, and constructing multi-scale detail feature maps for the two sets of images. Then, the local texture variance and feature response intensity of the corresponding regions in the two sets of detail feature maps are calculated. Based on the detail response intensity of the content image, the initial fusion weights of the corresponding regions are calculated. After global normalization in the 0-1 interval, the texture complementarity coefficients for image reconstruction are obtained. Then, the texture complementarity coefficients are downsampled to a size that perfectly matches each level of the multi-scale detail feature map using bilinear interpolation, obtaining the level-matched fusion weights. Based on the multi-scale structural features of the content image, a loss function constrained by multi-scale structural similarity is constructed, and pixel-wise weighted fusion is performed on the detail feature maps of each level. The gradient descent method is used to iteratively optimize the fusion process to eliminate structural bias. After iterative convergence, the fused features of each level are stitched together to generate an intermediate feature map with consistent texture of Hunan brocade. Finally, the pre-defined color style dictionary of Hunan brocade is constructed by training on the standard samples of classic Hunan brocade in the collection of intangible cultural heritage certified by the K-SVD dictionary learning algorithm. It contains the traditional standard color matching and weaving texture feature primitives of Hunan brocade. The intermediate feature map is mapped to the feature space of the dictionary. The optimal feature matching coefficient is solved by sparse coding. The color distribution and texture features of the pattern are adaptively corrected. Then, the feature is inversely transformed and mapped back to the image pixel space to obtain the optimized Hunan brocade pattern.

[0117] It should be noted that, in this application, detailed features refer to multi-scale feature maps that carry the details of the lines, weaving texture, and contours of Hunan brocade patterns; image reconstruction refers to the image processing process that completely preserves the pattern details of the original Hunan brocade to be optimized; texture complementarity coefficient refers to the fusion weight that controls the original pattern details of the content image and the target style texture of the style transfer image; multi-scale structural similarity constraint refers to the constraint rules that constrain the multi-scale structural integrity of the Hunan brocade pattern from global to local during the feature fusion process; feature-level fusion refers to the fusion of the content image within the deep feature domain extracted by the convolutional neural network. The process involves multi-scale, differentiated fusion of the original pattern details and the target style features of the style transfer image; the intermediate feature map refers to the fusion feature carrier that ensures the consistency of the texture across the entire surface of Hunan brocade; the Hunan brocade color style dictionary is a standardized feature dictionary constructed from standard samples of classic Hunan brocade in the collection that have been certified as intangible cultural heritage, containing traditional standard color matching and weaving texture features of Hunan brocade; adaptive reconstruction refers to the image reconstruction process that dynamically adjusts the color distribution and texture features of the pattern; the optimized Hunan brocade pattern refers to the final digital pattern that fully preserves the theme and pattern structure of the brocade to be optimized.

[0118] In this embodiment, reference Figure 4 As shown in the figure, this is a schematic diagram of the multi-scale feature extraction structure of Hunan brocade patterns. The figure illustrates the implementation process of decoupling and processing the content and style features of the brocade pattern. The original image layer is used as input for the Hunan brocade content image to be optimized; the hierarchical decomposition layer is used to perform multi-scale pyramid decomposition on the input image to obtain multi-resolution image components covering the global layout and local pattern details; the feature extraction layer is used to extract core detail features such as brocade pattern structure and weaving texture from the multi-scale components; the data processing layer is used to complete the style alignment and adaptive weighted fusion processing of the features; and the output result layer is used to output the fused features with texture consistency.

[0119] Therefore, in this application, image reconstruction is performed based on the detailed features of the style transfer image and the content image to generate an optimized Hunan brocade pattern. Specifically, determining the style guiding feature provides a directional style adjustment benchmark adapted to the multi-scale structural characteristics of Hunan brocade, thereby achieving hierarchical and spatially differentiated precise control over the full-page stylization process of the pattern. This style guiding feature anchors the core traditional style norms of Hunan brocade's intangible cultural heritage attributes, establishes a spatial correspondence between style features and the original pattern structure, and allows for setting differentiated style constraint strengths for the core pattern area and flat texture areas. While protecting the integrity of the original pattern structure, it ensures the adaptability and consistency of the full-page style features, thus providing a basis for the full-page style of the pattern. The precise adaptation and migration provides a stable and controllable basis; by determining the style transfer image, a full-width brocade image carrier that combines the integrity of the original pattern structure with the uniformity of the target style can be obtained, thus completing the entire link of Hunan brocade pattern from feature domain style constraints to pixel domain complete presentation. The style transfer image uses the target style information as a global constraint, completely continuing the core structural information such as the pattern layout and line outline of the original brocade, realizing a smooth transition and unified adaptation of the full-width traditional style, while repairing the details loss and texture breakage problems that occurred during the style transfer process, ensuring the consistency of the pattern style and the adaptability of the weaving process, and finally achieving the goal of completely preserving the original pattern structure and adapting and migrating the traditional style to the full-width image.

[0120] In summary, the technical solution adopted in this application can achieve full-frame style adaptation and transfer of patterns while completely preserving the original pattern structure of Hunan brocade.

[0121] Example 2: This application provides an intelligent optimization system for Hunan brocade patterns that integrates style transfer, referring to... Figure 5 As shown in the figure, this is a module structure diagram of an intelligent optimization system for Hunan brocade patterns that integrates style transfer, according to this embodiment of the present application. The optimization system includes:

[0122] The acquisition module 100 is used to acquire the content image and reference style image of the Hunan brocade pattern to be optimized;

[0123] The guided alignment module 200 is used to extract common style information of the content image and the reference style image, and to perform guided alignment on the common style information to obtain style guidance features for adjusting the pattern stylization.

[0124] The element filling module 300 is used to determine the splicing control conditions of the pattern when splicing features by using the common style information and the random noise features preset by the pattern stylization, generate target style information for style transfer of the pattern according to the splicing control conditions and the style guidance features, and then fill in the missing image areas during style transfer with the target style information to obtain a style-transferred image.

[0125] The image reconstruction module 400 is used to reconstruct the image based on the detailed features of the style transfer image and the content image, and generate an optimized Hunan brocade pattern.

[0126] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically-Erasable Programmable Read-Only Memory (EEPROM), compactdisc read-only memory (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium capable of carrying or storing data.

[0128] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

Claims

1. A method for intelligent optimization of Hunan brocade patterns by incorporating style transfer, characterized in that, The optimization method includes the following steps: Obtain the content image and reference style image of the Hunan brocade pattern to be optimized; Extract common style information from the content image and the reference style image, and perform guided alignment on the common style information to obtain style guidance features for adjusting pattern stylization; The splicing control conditions of the pattern during feature splicing are determined by the common style information and the random noise features preset by the pattern stylization. The target style information for style transfer of the pattern is generated according to the splicing control conditions and the style guidance features. Then, the missing image areas during style transfer are filled with elements by the target style information to obtain the style-transferred image. Image reconstruction is performed based on the detailed features of the style transfer image and the content image to generate an optimized Hunan brocade pattern.

2. The intelligent optimization method for Hunan brocade patterns integrating style transfer as described in claim 1, characterized in that, Extracting common style information between the content image and the reference style image specifically includes: The content image and the reference style image are decomposed into mutually independent structure-style distribution constraints by using a joint graph; Based on the structure-style distribution constraints, the common primitives of the content image and the reference style image are determined; The common primitives are injected into the structural phase and style magnitude corresponding to the content image and the reference style image, respectively, to obtain common style information.

3. The intelligent optimization method for Hunan brocade patterns integrating style transfer as described in claim 1, characterized in that, The style guidance features for adjusting pattern stylization are specifically obtained by guiding and aligning the common style information. Construct a dense field mapping relationship between the common style information and the target pattern space, and determine the style distribution weight of each pixel; The initial distribution vector in the spatiotemporal domain of the adjusted pattern stylization is determined based on the style distribution weights and the local structure tensor of the target pattern. The initial distribution vector is optimized for edge preservation to generate a spatially consistent alignment field; The spatial consistency alignment field is fused layer by layer with the multi-scale feature map of the target pattern to obtain style guidance features that adjust the pattern stylization.

4. The intelligent optimization method for Hunan brocade patterns integrating style transfer as described in claim 3, characterized in that, Determining the initial distribution vector in the spatiotemporal domain for adjusting pattern stylization based on the style distribution weights and the local structure tensor of the target pattern specifically includes: Based on the local structure tensor, feature decomposition is performed to extract the principal direction vector and anisotropic intensity coefficient of the target pattern at the pixel point; The principal direction vector is weighted and fused according to the style distribution weight and the anisotropy intensity coefficient to generate a local style guidance vector at each pixel. The local style-guided vector is regularized to obtain the initial distribution vector in the spatial domain when adjusting the pattern stylization.

5. The intelligent optimization method for Hunan brocade patterns based on style transfer as described in claim 1, characterized in that, The specific conditions for determining the splicing control of patterns during feature splicing, based on the common style information and the pre-defined random noise features of pattern stylization, include: A dual-modal hybrid sampling distribution is constructed in the joint feature space based on the probability distribution of the common style information and the random noise features; The spatial adaptive control coefficients of each feature channel are generated during feature stitching using the dual-modal mixed sampling distribution and the preset learnable stitching weight mapping function. Based on the spatial adaptive control coefficients and the local correlation metric between feature maps, the stitching control conditions for the pattern during feature stitching are determined.

6. The intelligent optimization method for Hunan brocade patterns integrating style transfer as described in claim 1, characterized in that, The specific methods for generating target style information for style transfer of patterns based on the splicing control conditions and the style guidance features include: Construct a multi-scale distribution mapping relationship in the joint feature space based on the splicing control conditions and the style guidance features; Based on the multi-scale distribution mapping relationship and the spatial location encoding of the target pattern, the dynamic modulation coefficients of style information in each spatial region are determined; The style guidance features are spatially adaptively fused using the dynamic modulation coefficients to generate a multi-channel feature map of preliminary target style information. The multi-channel feature map is regularized iteratively based on the boundary continuity constraint in the splicing control conditions to obtain the target style information when performing style transfer on the pattern.

7. The intelligent optimization method for Hunan brocade patterns integrating style transfer as described in claim 6, characterized in that, Determining the dynamic modulation coefficients of style information in each spatial region based on the multi-scale distribution mapping relationship and the spatial location encoding of the target pattern specifically includes: Based on the multi-scale distribution mapping relationship and the spatial location encoding, a conditional probability density distribution function is constructed in the joint feature space; The preliminary modulation weights of the regional style response are calculated by integrating the conditional probability density distribution function over the spatial regions of the target pattern. Adaptive regularization is performed based on the initial modulation weights and neighborhood space continuity constraints to obtain the dynamic modulation coefficients of style information in each spatial region.

8. The intelligent optimization method for Hunan brocade patterns integrating style transfer as described in claim 1, characterized in that, The missing image regions during style transfer are filled with elements using the target style information to obtain the style-transferred image, specifically including: Based on the target style information and the known pixels in the neighborhood of the missing image region during style transfer, determine the style guidance distribution of each pixel in the missing region; The set of confidence elements within the missing region is determined by the style-guided distribution; The set of confidence elements is subjected to style consistency processing to obtain a style transfer image.

9. The intelligent optimization method for Hunan brocade patterns integrating style transfer as described in claim 1, characterized in that, Image reconstruction based on the detailed features of the style-transferred image and the content image to generate an optimized Hunan brocade pattern specifically includes: The texture complementarity coefficients for image reconstruction are determined based on the detail feature maps of the style transfer image and the content image. Feature-level fusion is performed using the texture complementarity coefficient and multi-scale structural similarity constraints to generate an intermediate feature map with consistent texture of Hunan brocade. The optimized Hunan brocade pattern is obtained by adaptive reconstruction based on the intermediate feature map and the preset Hunan brocade color style dictionary.

10. A system for intelligent optimization of Hunan brocade patterns incorporating style transfer, used to execute the method for intelligent optimization of Hunan brocade patterns incorporating style transfer as described in any one of claims 1 to 9, characterized in that, The optimization system includes: The acquisition module is used to acquire the content image and reference style image of the Hunan brocade pattern to be optimized; The guided alignment module is used to extract common style information between the content image and the reference style image, and to perform guided alignment on the common style information to obtain style guidance features for adjusting the pattern stylization. The element filling module is used to determine the splicing control conditions of the pattern when splicing features by using the common style information and the random noise features preset by the pattern stylization, generate target style information for style transfer of the pattern according to the splicing control conditions and the style guidance features, and then fill in the missing image areas during style transfer with the target style information to obtain the style transfer image. The image reconstruction module is used to reconstruct the image based on the detailed features of the style transfer image and the content image, and generate an optimized Hunan brocade pattern.

Images